MNKNDLTD/kenya-election-companion

4-persona AI Election Companion that fights misinformation and drives agency in civic + democratic participation README.md # Kenya 2027 Election Companion — 4-Persona Demo **Mozilla Foundation AIxD Incubator Grant Submission** **Organization:** MNKND Limited **Grant Amount:** $50,000 USD **Duration:** 14 months (Jul 2026 – Aug 2027) --- ## Overview Kenya 2027 Election Companion is a WhatsApp-native, AI-powered civic infrastructure platform that fact-checks election rumors in under 10 seconds, makes party manifestos conversationally accessible in English/Swahili/Sheng, tracks live verified results on election day, and surfaces emerging misinformation trends to civil-society partners 12–24 hours before they peak. **Target:** 22 million Kenyan voters, especially first-time, rural, and women voters most targeted by misinformation. **Partnership:** Built alongside Royal Media Services (Kenya's largest media group: Citizen TV, Radio Citizen, 11 vernacular stations), Uraia Trust, and Code for Africa. --- ## 4-Persona Architecture This demo implements the complete 4-persona system described in the Mozilla grant application: ### **1. VOTER (WhatsApp Interface)** - Fact-check rumors (submit claim, get verdict in 80%, route to human if 80%, auto-publish verdict → If AI confidence <80%, route to Newsroom Editor claim queue Newsroom Editor publishes verdict → Update verdicts table with editor signature → Voter receives WhatsApp notification with verdict → Civil Society dashboard sees new claim in misinfo_clusters → Public API exposes new row in /api/v1/claims (anonymized) Civil Society activates intervention → Log intervention in misinfo_clusters.intervention_status → SMS blast sent to voters in affected county → RMS contacted for on-air correction ``` --- ## Mozilla Grant Requirements Met ### **C.1 Technical Architecture ✅** - Data-first architecture (RMS archive, IEBC docs, manifestos) - 4-layer verification (RMS → external sources → correspondents → editorial) - Multilingual normalizer (English/Swahili/Sheng code-switching) - RAG with source citations (not model memory) …